Enhanced energy demand management in electric distribution networks using lstm-xgboost model

Accurately forecasting electricity consumption is essential for improving the reliability, operational efficiency, and resilience of energy systems, particularly given the rising integration of renewables and the advancing complexity of smart grids. This research introduces an innovative multilayered model for short-term electric load forecasting that combines Long Short-Term Memory (LSTM) networks with Extreme Gradient Boosting (XGBoost) to enhance accuracy and dependability. The approach consists of three main steps: data cleaning and feature extraction, development of separate LSTM and XGBoost models, and combining them into a unified hybrid architecture. The Elia Grid dataset from Belgium was used in this study, containing high-resolution load data for 2022 captured at 15-minute intervals. The hybrid model leveraged the LSTM's strength in learning sequential dependencies, while XGBoost contributed by capturing non-linear residual patterns and extracting feature importance. The proposed model underwent extensive testing to evaluate its performance against independent LSTM and XGBoost models. The hybrid model achieved its best results with a Root Mean Square Error (RMSE) of 106.54 MW and Mean Absolute Percentage Error (MAPE) of 1.18% and Coefficient of Determination (R²) of 0.994. The sensitivity analysis showed that increasing the look-back window size improved model performance but the attention mechanisms did not enhance accuracy so they were removed from the final design. In addition to outperforming traditional models, the proposed framework demonstrated strong generalizability, scalability, and interpretability, making it suitable for real-time energy management systems. The model was benchmarked against recent classical and deep learning (DL) models and showed competitive or superior results across multiple datasets. This work contributes to the design of AI-based applications in smart grid management and enriches the literature on hybrid DL techniques for time series forecasting. Future research will explore expanding the model to multi-regional datasets, incorporating weather and socioeconomic variables, and deploying the system in real-world grid control environments.